基于分频段ABAP谱减法的鸟叫声分类研究

Study of the Bird Sound Classification Based on ABAP Multi-band Spectral Subtraction

摘要: 针对带噪鸟叫声,用自适应Bartlett平均周期图(ABAP)完成噪声估计,进行分频段谱减去噪后,提取经过二维离散余弦变换的Mel频率倒谱系数动态声音特征(TDMFCC).最后,利用支持向量机(SVM)分别结合MFCC、TDMFCC以及经过分频段ABAP谱减法降噪后的MFCC和TDMFCC对30种鸟叫声进行不同背景环境和信噪比情况下的对比实验.结果表明,分频段ABAP谱减法降噪后提取TDMFCC结合SVM的方法可以取得较好的分类效果,适用于噪声环境下的鸟叫声分类.

Abstract: To the bird sounds mixed with noises,the estimation of noise is done by using the adaptive Bartlett averaging periodogram (ABAP), and noise reduction is made by using multi-band spectral subtraction, then the dynamic feature of Mel-frequency spectrum coefficients by two-dimensional discrete cosine transform (TDMFCC)was extracted. The comparison experiments of 30 bird sounds classification in different environments under different SNRs were constructed based on the combination of SVM classifier and different features, namely MFCC, TDMFCC and MFCC, TDMFCC through ABAP multi-band spectral subtraction respectively. The results showed that the method of TDMFCC through ABAP multi-band spectral subtraction combined SVM can achieve better classification effect, which is very suitable for the bird sound classification in noisy environments